中国人民公安大学

Multimodal emotion recognition (MER) aims to classify emotions by integrating heterogeneous information sources, playing an increasingly pivotal role in various domains. However, MER, due to its information heterogeneity, usually suffers from inharmonious feature fusion, particularly in modeling dynamic cross-modal interactions and intrinsic speaker attributes, resulting in low recognition accuracy. Furthermore, many existing methods rely on complex networks to handle multimodal information, possibly incurring high computational overhead. To address these issues, we propose a novel and efficient Gated Cross-Attention with Speaker-injected Mamba for multimodal emotion recognition, named GCA-SiM, which on the one hand, captures the complementary semantic information between diverse modalities and on the other hand, achieves efficient modeling of long-term dependencies in speaker context. Specifically, we design an adaptive gated fusion block (AGFBlock) that employs a symmetric gating mechanism to adaptively filter intra-modal noise and recalibrate inter-modal contributions, capturing complementary semantic cues through a cross-attention strategy. To strike a balance between efficiency and performance, we further develop the speaker-shift Mamba (S2Mamba) that incorporates speaker identities and speaker-shift cues into the state space model (SSM) via an additive injection mechanism, enabling efficient modeling of long-range conversational dependencies. Additionally, a focal loss function is employed to mitigate the impact of emotion class imbalance, focusing the training on hard-to-classify samples. Extensive experiments on the benchmark datasets IEMOCAP and MELD demonstrate that GCA-SiM achieves competitive performance while significantly reducing memory consumption and training time. Notably, GCA-SiM consumes only 43.26 MB of memory and 2.50 s per epoch on the MELD datasets, respectively, far less than other compared methods, indicating its high efficiency.

In event-driven cyberbullying, hostile attention can rapidly converge on identifiable targets, exposing them to repeated collective attacks. Existing research has advanced harmful-content detection and the analysis of online aggression, but offers limited insight into how such attacks become concentrated on particular targets. To address this gap, this study constructs event-level trajectories of attacks against specific targets and introduces Targeted Aggression Propagation, a mechanistic simulation framework for modeling how attack centers emerge and evolve. The framework links overall event dynamics with the uneven accumulation of attacks across targets. We evaluate this approach on five cyberbullying cases using public Sina Weibo posts and first-level comments. Empirical evaluation shows that the framework reproduces observed multi-wave comment dynamics and characterizes changes in target dominance. Across the five cases, attack centers take distinct forms, including stable single-target dominance, two-target concentration, and competitive allocation. Scenario simulations further clarify how different mechanisms shape these forms. Post-driven stimulation, comment-driven reinforcement, and cooling mainly regulate the scale and persistence of targeted exposure, but their effects weaken after the main activation period. By contrast, mechanisms at the target layer shape how exposure becomes concentrated, stabilized, or redistributed under similar levels of overall aggression. These findings show that cyberbullying events should be examined not only by harmful-content occurrence or aggregate aggression, but also through the targets repeatedly exposed to collective attacks.

This study examines Japan’s policy adjustments between 2023 and 2025 in anticipation of Donald Trump’s return to the White House. It argues that these adjustments are best understood as Japan’s attempt to navigate the enduring alliance dilemma—the perennial tension between fear of abandonment by its senior partner and fear of entrapment in its conflicts. The shocks of Trump’s first term (2017–2021)—trade pressure, security uncertainty, and unilateralism—provided a powerful historical catalyst, sharply intensifying Tokyo’s perception of these twin risks. In response, Japan adopted a composite strategy blending proactive engagement with prudent risk hedging across high-level diplomacy, economic diversification, defense posture, and multilateral institution-building. The paper finds that Japan’s behavior resembles a dynamic risk mitigation mechanism: activated by historical experience but fundamentally shaped by the structural logic of the alliance dilemma, rather than a fully developed grand strategy. Its effectiveness remains constrained by domestic politics, U.S.–China competition, and the future trajectory of U.S. policy. The case illustrates both the possibilities and the inherent limits of hedging as a middle-power strategy under conditions of hegemonic uncertainty.

Despite the success of test-time scaling, Large Reasoning Models (LRMs) frequently encounter repetitive loops that lead to computational waste and inference failure. In this paper, we identify a distinct failure mode termed Circular Reasoning. Unlike traditional model degeneration, this phenomenon manifests as a self-reinforcing trap where generated content acts as a logical premise for its own recurrence, compelling the reiteration of preceding text. To systematically analyze this phenomenon, we introduce LoopBench, a dataset designed to capture two distinct loop typologies: numerical loops and statement loops. Mechanistically, we characterize circular reasoning as a state collapse exhibiting distinct boundaries, where semantic repetition precedes textual repetition. We reveal that reasoning impasses trigger the loop onset, which subsequently persists as an inescapable cycle driven by a self-reinforcing V-shaped attention mechanism. Guided by these findings, we employ the Cumulative Sum (CUSUM) algorithm to capture these precursors for early loop prediction. Experiments across diverse LRMs validate its accuracy and elucidate the stability of long-chain reasoning.

Low-Rank Adaptation (LoRA) has emerged as a leading technique for efficiently fine-tuning text-to-image diffusion models, and its widespread adoption on open-source platforms has fostered a vibrant culture of model sharing and customization. However, the same modular and plug-and-play flexibility that makes LoRA appealing also introduces a broader attack sur-face. To highlight this risk, we propose Masquerade-LoRA (MasqLoRA), the first backdoor attack that leverages the LoRA mechanism to stealthily inject malicious behavior into text-to-image diffusion models. MasqLoRA operates by freezing the base model parameters and updating only the low-rank adapter weights using a small number of “trigger word-target image” pairs. This enables the attacker to train a standalone backdoor LoRA module that embeds a hidden cross-modal mapping: when the module is loaded and a specific textual trigger is provided, the model produces a predefined visual output; other-wise, it behaves indistinguishably from the clean model, ensuring the stealthiness of the attack. Experimental results demonstrate that MasqLoRA can be trained with minimal resource overhead and achieves a high attack success rate of 99.8%. MasqLoRA reveals a severe and unique threat in the AI supply chain, underscoring the urgent need for dedicated defense mechanisms for the LoRA-centric sharing ecosystem.